Midjourney's first hardware product is a full-body ultrasonic CT scanner
Midjourney unveiled a full-body ultrasonic CT scanner as its first hardware product, G7 leaders demanded AI sovereignty guarantees following the Anthropic export ban, and Hugging Face launched an open protocol for agents to discover capabilities at runtime. DeepSeek added vision capabilities to its chat product, world model startup Odyssey raised $310 million at a $1.45 billion valuation, and today’s ArXiv batch introduced conformal risk control for tool-using agents and a benchmark showing current agents fail at long-horizon strategic tasks.
Regulatory & Policy #
G7 leaders demand AI sovereignty guarantees after Anthropic export ban #
TechCrunch
At the G7 summit, French President Macron warned that the US ability to “turn off the switch” on AI access threatens allied economies, while Indian PM Modi argued democratic nations need unrestricted access to top AI models for critical infrastructure. The Anthropic Mythos 5 / Fable 5 export ban days earlier transformed this from theoretical concern to demonstrated vulnerability, prompting leaders to propose a “trusted partners” scheme for maintaining AI access that bypasses unilateral US restrictions. For organizations dependent on American AI providers, the G7 response signals that sovereign AI alternatives and multi-provider strategies are becoming a geopolitical priority, not just a procurement preference.
US holds off blacklisting DeepSeek while adding 100+ firms to security list #
Reuters / Hacker News (465 points)
The US government declined to add DeepSeek to its entity list even as it designated over 100 other Chinese firms as security risks, signaling that the administration views open-source Chinese AI models differently from hardware or defense companies. The decision comes amid ongoing debate about whether restricting access to widely distributed open-weight models is technically enforceable or strategically counterproductive. For teams using or evaluating DeepSeek models, the reprieve reduces near-term regulatory risk but does not resolve the underlying policy uncertainty.
Model Releases #
DeepSeek introduces vision capabilities #
DeepSeek / Hacker News (170 points)
DeepSeek enabled vision and image understanding in its consumer chat product, extending the V4 model family with capabilities for diagram parsing, screenshot-to-code generation, and document extraction from images. The addition makes DeepSeek’s chat interface multimodal without requiring a separate model, following the pattern set by OpenAI and Anthropic of integrating vision as a core modality rather than a specialized endpoint. For developers evaluating multimodal options, DeepSeek’s open-weight approach means these vision capabilities will likely become available for local deployment.
Developer Tools #
Agentic Resource Discovery: Let agents search for capabilities at runtime #
Hugging Face
Hugging Face launched Agentic Resource Discovery (ARD), a draft open specification developed with Microsoft, Google, and GoDaddy that defines a standard way for agents to dynamically find tools, skills, and other agents across federated registries using natural language queries. The spec includes a static manifest format (ai-catalog.json) for publishers and a dynamic registry API with a POST /search endpoint, shifting from manually installed tool catalogs to intent-based search outside the LLM context window. For teams building agents that need ad-hoc capabilities, ARD addresses a real scaling bottleneck: hardcoding tool configurations or dumping everything into context both break as the tool ecosystem grows.
New Amazon Bedrock AgentCore capabilities for knowledge, observability, and security #
AWS Machine Learning Blog
AWS announced new Bedrock AgentCore capabilities spanning three layers: organizational knowledge via managed knowledge bases with agentic retrieval across SharePoint, Confluence, and S3; world knowledge through live web search within AWS security boundaries; and paid knowledge enabling agents to access licensed content while compensating providers. The release also includes failure pattern detection that ranks silent behavioral failures by impact and recommends prompt and tool improvements. For teams deploying agents in enterprise settings, the combination of multi-source knowledge retrieval with continuous observability addresses the gap between prototype agents that work on curated data and production agents that need to operate across messy real-world information sources.
NVIDIA AI coding agents autonomously direct robot training #
Ars Technica / NVIDIA
NVIDIA demonstrated a system where teams of AI coding agents autonomously write and iterate on robot training code, teaching robots to perform tasks like installing GPUs and cutting zip ties without human-authored training scripts. The approach uses coding agents as an intermediary between high-level task descriptions and low-level robot control policies, with the agents generating, testing, and refining training code through self-improvement loops. This collapses the bottleneck between specifying what a robot should do and producing the training pipeline that teaches it, a step that currently requires specialized robotics engineers.
Research & Papers #
CEO-Bench: Can Agents Play the Long Game? #
arxiv:cs.CL
CEO-Bench tests capabilities that current agent benchmarks largely ignore: navigating long horizons amid uncertainty, acquiring information in noisy environments, adapting to a changing world, and orchestrating multiple moving parts toward a coherent goal. The benchmark reveals that while agents handle isolated short-horizon tasks well, they fail at the combination of skills required for sustained strategic execution. For teams building agentic workflows, this quantifies the gap between task-level competence and the kind of multi-step, adaptive planning that real-world deployment demands.
ToolChain-CRC: Conformal Risk Control for Agentic AI Under Retrieval and Tool-Use Drift #
arxiv:cs.LG
ToolChain-CRC introduces the first conformal risk-control method for retrieval-augmented and tool-using agents, catching failures that are invisible from the final answer alone – cases where retrieval was weak, a tool output was wrong, or an earlier step was unsupported. The method provides statistical guarantees on multi-step agent reliability under distribution drift, addressing the production reality that retrieval quality and tool behavior shift over time. For teams deploying agents against live data, this fills a critical gap: existing evaluation methods test agents against static data, but production agents face non-stationary tool and retrieval environments.
VISUALSKILL: Multimodal Skills for Computer-Use Agents #
arxiv:cs.CL
VISUALSKILL proposes hierarchical multimodal skill libraries for computer-use agents, arguing that text-only skill representations miss the fundamentally visual nature of GUI interaction. The framework organizes skills as a central index over per-topic files tailored to each target application, enabling agents to reuse learned interaction patterns across sessions. Computer-use agents currently struggle on long-horizon tasks partly because they rediscover UI patterns from scratch each time; structured visual skills could reduce this redundancy.
Towards an Agent-First Web: Redesigning the Web for AI Agents #
Hugging Face Daily Papers
This paper argues that the web’s three-decade-old assumption – that the primary consumer of content is a human – is now invalid, and that the resulting mismatch causes blanket agent blocking, CAPTCHA exclusion, and economic models that treat agent traffic as parasitic. The authors propose redesigning web infrastructure around agent access patterns as a first-class concern rather than an adversarial edge case. The argument is structurally important: as agents become primary web consumers, the current friction is not just an inconvenience but a systematic barrier to the agent-mediated web that both HuggingFace’s ARD spec and every RAG pipeline implicitly depend on.
GateMem: Benchmarking Memory Governance in Multi-Principal Shared-Memory Agents #
arxiv:cs.CL / arxiv:cs.LG
GateMem introduces a benchmark for agents that share memory across multiple principals – the hospital, workplace, and campus deployments where different users write to a common memory pool under different roles and scopes. Existing memory benchmarks assume single-user settings, leaving the governance dimension (who can read what, under which role) entirely unevaluated. For teams building shared AI assistants, this highlights that memory quality in multi-user settings requires access control and scoping, not just better recall.
Data Intelligence Agents: Autonomous Coding Agents for Enterprise Data Integration #
Hugging Face Daily Papers
Data Intelligence Agents compresses the multi-handoff workflow between data owners, engineers, and analysts into three autonomous coding agents that discover, structure, and query enterprise data. Rather than emitting text, the agents generate, execute, validate, and repair concrete code artifacts – an architecture choice that trades prompt-engineering fragility for deterministic execution. The paper addresses a real enterprise bottleneck where data integration is delayed not by technical complexity but by repeated human handoffs that lose context at each step.
OpenAI LifeSciBench: Expert-Authored Benchmark for Life Science AI #
OpenAI
OpenAI introduced LifeSciBench, an expert-authored benchmark for evaluating how AI systems handle real-world life science research tasks and decisions, designed to test scientific reasoning rather than pattern-matched answers from training data. The benchmark addresses a gap in AI evaluation for science applications, where existing benchmarks often test knowledge retrieval rather than the judgment calls that actual researchers face. For teams deploying AI in life science workflows, this provides a more realistic evaluation signal than general-purpose reasoning benchmarks.
Security #
“Dangerous” AI models with advanced hacking capabilities will soon be the norm #
Ars Technica
Ars Technica reports that AI models capable of autonomous hacking, vulnerability discovery, and exploit generation are approaching commodity status, with the capability gap between frontier and open-weight models narrowing faster than defensive tooling can adapt. The article argues that security-focused capability restrictions on individual models are increasingly futile as the underlying techniques proliferate. For security teams, this reinforces the shift from “prevent AI from having offensive capabilities” to “assume offensive AI capabilities exist and build defenses accordingly.”
Funding & Business #
Odyssey raises $310M at $1.45B valuation for world models #
TechCrunch
Odyssey, a startup building AI world models that simulate physical environments with accurate physics, raised $310 million in a Series B led by Natural Capital with participation from Amazon, AMD Ventures, and GV, reaching a $1.45 billion valuation. The company deploys people with cameras to collect real-world data, then generates interactive simulations for applications from video games to robotics. World models represent a distinct AI paradigm from language models – simulating physical dynamics rather than text patterns – and the investment signals that the physical-AI data and simulation layer is attracting infrastructure-scale capital.
Pramaana Labs raises $27M to bring formal verification to AI #
TechCrunch
Pramaana Labs raised $27 million in seed funding led by Khosla Ventures to build formal verification layers on top of LLMs, using the LEAN programming language to mathematically prove that AI outputs are correct in high-stakes domains like law, drug discovery, and tax preparation. The approach codifies domain rules into deterministic verification systems so that AI-generated outputs can be checked against provable constraints rather than relying on probabilistic confidence. For teams deploying AI in regulated verticals where errors carry legal or safety consequences, formal verification offers a structural alternative to the “add more guardrails” approach.
XDOF raises $70M for robot training data infrastructure #
TechCrunch
XDOF emerged from stealth with $70 million in funding to build data collection pipelines, tooling, and annotation systems for robot training, working with approximately 20 customers including several frontier AI labs. Physical AI faces a data problem that language models solved through web scraping: robots require high-fidelity data capturing physical interactions, which barely exists at scale. XDOF is betting that robot training data is an infrastructure layer, not a one-off collection effort – a thesis that parallels how Scale AI built a business around the data annotation bottleneck for language and vision models.
DeepL acquires Mixhalo for live-event audio streaming and translation #
TechCrunch
German translation company DeepL acquired Mixhalo, a real-time audio platform for live events founded by Incubus guitarist Mike Einziger, to expand into live audio translation for sports and entertainment venues. DeepL is opening a San Francisco office as part of the deal. The acquisition extends AI translation from text into the real-time audio domain, where latency constraints and ambient noise create challenges that document translation models do not face.
Open Source #
Local Qwen is not a worse Opus – it is a different tool #
Alex Ellis / Hacker News (235 points)
Alex Ellis argues through practical experience that local 27B Qwen models should not be evaluated as inferior versions of frontier models but as fundamentally different tools suited to bounded, well-defined tasks. Local models excel at explaining existing codebases and specialized work (customer diagnostics, maintenance, testing) but fail at long-horizon unsupervised work, entering infinite loops on complex tasks. The post identifies specific practical boundaries: local models need detailed instructions (AGENTS.md format), careful quantization settings, and should never be left unsupervised on distributed systems code. The article complements the broader local-model viability discussion with concrete failure modes and effective usage patterns.
Adam: Open-Source AI CAD (YC W25) #
Adam / Hacker News
YC W25 company Adam launched CADAM, an open-source platform for AI-generated mechanical CAD, built on the premise that text-to-code-to-CAD is the right generation paradigm for mechanical design. The React app with Supabase backend generates CAD through code rather than direct geometry manipulation, paralleling how AI coding tools generate software through text rather than visual editors. For teams working at the intersection of AI and physical product design, this is an early signal that the text-to-code pattern is extending beyond software into mechanical engineering.
Infrastructure #
France advances Europe’s AI future with NVIDIA technologies #
NVIDIA Blog
One year after NVIDIA GTC Paris laid out plans for French AI infrastructure, the investments are coming online: AI factories are operational, national compute capacity is deployed, and startups are building applications on the infrastructure. AI agents are running in production across the French ecosystem, with the country positioning itself as Europe’s AI hub through a combination of public compute investment and NVIDIA partnership. The progress demonstrates that national AI compute strategies can move from announcement to deployment within a year when backed by sufficient capital and an established hardware partner.
Other #
Midjourney Medical: full-body ultrasonic CT scanner #
Midjourney / Engadget / Hacker News (797 points)
Midjourney announced its first hardware product: a full-body ultrasonic CT scanner using 500,000 transducers that captures a complete body image in 60 seconds with no radiation, developed in partnership with Butterfly Network’s ultrasound-on-chip technology. The company plans a San Francisco “Spa” location in 2027, with custom silicon for improved image quality in 2028, and an ambition of 50,000 scanners worldwide by 2031. The pivot from AI image generation to medical imaging hardware is surprising in scope but leverages Midjourney’s core competency in signal-to-image processing; whether the company can navigate medical device regulation as effectively as it navigated the generative AI market remains the open question.
OpenAI AI chemist improves a challenging reaction in medicinal chemistry #
OpenAI / Molecule.one
OpenAI and Molecule.one demonstrated a near-autonomous AI chemist using GPT-5.4 that improved a key drug-making reaction in medicinal chemistry, with the system operating with minimal human intervention to optimize reaction conditions. The work extends the pattern of frontier models being applied to scientific domains where the search space is too large for manual experimentation but structured enough for AI-guided exploration.
Anthropic becomes first AI startup to join Frontier carbon removal coalition #
TechCrunch / Frontier
Anthropic joined the Frontier coalition, a carbon removal purchasing group founded by Stripe, Google, and Shopify, which received $915 million in new pledges bringing total commitments to $1.8 billion. Frontier plans to shift toward fewer but larger projects with 8-10 year contracts targeting gigaton-scale removal. As AI compute energy consumption draws increasing scrutiny, Anthropic’s membership signals that frontier labs are beginning to address their energy footprint through structured purchasing commitments rather than ad-hoc offsets.
Only 16% of Americans think AI will have a positive impact on society #
TechCrunch / Pew Research Center
A Pew Research Center study found that only 16% of Americans believe AI will have a positive impact on society, a significant gap between public sentiment and the industry’s self-assessment. The finding matters for AI companies because sustained negative public perception eventually translates into regulatory pressure, user resistance, and talent pipeline challenges regardless of technical capabilities.
Threads to Watch #
Physical AI is becoming an investable infrastructure category. Odyssey raised $310 million for world models that simulate physical environments, XDOF raised $70 million for robot training data pipelines, Midjourney pivoted to medical imaging hardware, and NVIDIA demonstrated coding agents that autonomously write robot training code. The common thread: physical AI requires purpose-built data, simulation, and tooling layers that do not exist yet, and capital is flowing to build them.
Agent interoperability standards are emerging. Hugging Face’s Agentic Resource Discovery spec, AWS Bedrock AgentCore’s multi-source knowledge layer, and today’s ArXiv papers on memory governance and risk control for tool-using agents all point toward the same need: agents operating in production cannot be self-contained – they need standardized ways to discover capabilities, access knowledge across boundaries, and provide reliability guarantees. The shift from “build a better agent” to “build the infrastructure agents need” mirrors the early web’s evolution from better browsers to HTTP standards.
The Anthropic export control ripple effects keep expanding. The G7 “trusted partners” proposal, the US decision to hold off on blacklisting DeepSeek, and the Pew finding that only 16% of Americans view AI positively all connect to the same underlying tension: governments want to control AI access for security reasons while the technical and economic reality makes unilateral control increasingly difficult to sustain without damaging allied relationships and domestic innovation.
Sources Unavailable Today #
These sources could not be fetched today. Links point to their homepages so you can check them directly.
- Weights & Biases: Fully Connected — scrape: content not extractable
- Stanford HAI — scrape: content truncated